Pipeline runtime

Graph, Run, options, reports, errors. More...

Classes Index

struct	DebugOptions
	Tunables for debug-mode pipeline dumps. More...
struct	DebugOutput
	One captured pipeline output (single-shot dump result). More...
struct	DebugStream
	Streaming iterator over debug-mode pipeline outputs. More...
struct	Box
	One axis-aligned detection produced by BoxDecode. More...
struct	BoxDecodeResult
	Parsed BoxDecode output paired with its raw byte buffer. More...
struct	FormatSpec
	Thin wrapper around FormatTag with implicit string conversions. More...
class	RtspServerHandle
	Live handle for a Graph running in RTSP server mode. More...
class	Graph
	The assembly stage — turns a list of Nodes into a runnable, deterministic pipeline. More...
struct	GroupMeta
	Per-fragment metadata captured during build. More...
struct	RtspServerOptions
	Options for Graph::run_rtsp() — controls the RTSP server's mount point and ports. More...
struct	AdvancedExecutionOptions
	Intent-named, jargon-free execution controls (preferred over the raw legacy fields). More...
struct	GraphOptions
	Per-Graph construction options. More...
struct	OutputTensorOptions
	Options for Graph::add_output_tensor() — the tensor-friendly output helper. More...
struct	Sample
	Typed payload returned by Run::pull() and consumed by Run::push(). More...
struct	RunOptions
	Per-Run runtime options. More...
class	Run
	Live pipeline handle: push inputs in, pull outputs out. More...
struct	BoxDecodeOptions
	Options driving a standalone BoxDecode invocation. More...

Typedefs Index

using	SampleAttributes = std::map< std::string, std::string >
	Per-frame user attributes carried alongside a Sample. More...

Enumerations Index

enum class	BoxDecodeType : std::int32_t { ... }
	Decode families accepted by the BoxDecode backend. More...
enum class	BoxDecodeTypeOption : std::int32_t { ... }
	Tensor packing/layout option within a decode family. More...
enum class	BoxDecodeSourceStorage : std::int32_t { ... }
	Source byte layout of the head tensors feeding a BoxDecode stage. More...
enum class	FormatTag { ... }
	Identifies a media or tensor payload format. More...
enum class	RunMode { ... }
	Internal timing mode a Run operates in. More...
enum class	SampleKind { ... }
	What kind of payload a Sample carries. More...
enum class	PullStatus { ... }
	Result status of Run::pull(). More...
enum class	OverflowPolicy { ... }
	What push() does when the input queue is full. More...
enum class	RunPreset { ... }
	Convenience preset bundles for RunOptions. More...
enum class	OutputMemory { ... }
	How output Tensors relate to the underlying GStreamer buffers. More...

Description

Graph, Run, options, reports, errors.

Typedefs

SampleAttributes

using simaai::neat::SampleAttributes = typedef std::map<std::string, std::string>

Per-frame user attributes carried alongside a Sample.

A plain string-to-string map. Values are opaque to the framework: it copies them, keeps them associated with their frame, and clears them when a destination buffer is reused. It never parses, merges, or reinterprets them.

Keys are normalized to ASCII lowercase by producers that capture them from a transport (see MultipartHeaderCaptureOptions). Ordering is lexicographic by key. Keys and values must be valid UTF-8 and may not contain embedded NUL bytes because the GStreamer string representation cannot preserve other encodings.

See Also

Sample::attributes

See Also

MultipartHeaderCaptureOptions

Definition at line 513 of file GraphOptions.h.

Enumerations

BoxDecodeSourceStorage

enum class simaai::neat::BoxDecodeSourceStorage : std::int32_t

strong

Source byte layout of the head tensors feeding a BoxDecode stage.

Enumeration values

DenseHwc	Dense, physically contiguous HWC heads (e.g. a Cast/dequant output) (= 0)
PackedCBlock	Channel-block packed / tessellated heads (e.g. a cblock detess output) (= 1)

For model-pack (Model-constructed) graphs this is resolved authoritatively from the model archive and must NOT be set by hand. For hand-built graphs that decode without a model pack (e.g. a Cast/detess stage stitched directly into SimaBoxDecode), the upstream contract does not carry the packing flags, so the value cannot be inferred and must be specified explicitly via the SimaBoxDecode constructor. When it is neither supplied by a model pack nor specified, contract compilation fails fast.

Definition at line 117 of file BoxDecodeType.h.

BoxDecodeType

enum class simaai::neat::BoxDecodeType : std::int32_t

strong

Decode families accepted by the BoxDecode backend.

Enumeration values

Unspecified	Sentinel: no decode family selected (fails fast at runtime) (= 0)
Yolo	Generic YOLO family token (= 1)
YoloV5	YOLOv5 detection (= 2)
YoloV5Seg	YOLOv5 segmentation (= 3)
YoloV7	YOLOv7 detection (= 4)
YoloV7Seg	YOLOv7 segmentation (= 5)
YoloV8	YOLOv8 detection (= 6)
YoloV8Seg	YOLOv8 segmentation (= 7)
YoloV8Pose	YOLOv8 pose-estimation heads (= 8)
YoloV9	YOLOv9 detection (= 9)
YoloV9Seg	YOLOv9 segmentation (= 10)
YoloV10	YOLOv10 detection (= 11)
YoloV10Seg	YOLOv10 segmentation (= 12)
Detr	DETR-style transformer detection (= 13)
EffDet	EfficientDet detection (= 14)
RcnnStage1	Region-proposal stage of two-stage R-CNN models (= 15)
Centernet	CenterNet keypoint-style detection (= 16)
YoloV26	YOLO26 detection (raw l/t/r/b distance heads) (= 17)
YoloV26Pose	YOLO26 pose-estimation heads (= 18)
YoloV26Seg	YOLO26 segmentation heads (= 19)
YoloV6	YOLOv6 raw l/t/r/b distance heads (= 20)
YoloX	YOLOX raw xywh heads with separate objectness and class logits (= 21)
Ssd	SSD family token, resolved internally to an exact supported prepared signature (= 22)
SuperPoint	SuperPoint detector-logit and descriptor-grid postprocessing (= 23)

Unspecified is an internal unset sentinel and must fail fast before runtime decode. Most YOLO-family variants share the same class-inference contract in genericboxdecode_v2:

• decoupled heads: repeated class-depth tensors, class depth > 4
• packed heads: depth = 3 * (num_classes + 5), consistent across heads YoloV26 uses decoupled 4-channel raw l/t/r/b bbox heads paired with class heads. YoloV26Seg uses the same raw l/t/r/b bbox heads, class-score heads, 32-channel mask-coefficient heads, and a trailing mask prototype. YoloV26Pose uses the same raw l/t/r/b bbox heads, 1-channel pose scores, and 51-channel keypoint heads. Ssd selects one of four prepared SSD contracts (not a generic SSD decoder), resolved from the complete ordered box-decode head geometry at compile time:
  • SSD300: feats {38,19,10,5,3,1}, priors-per-cell {4,6,6,6,4,4}, 300x300, softmax.
  • SSD-Mobile-300: feats {19,10,5,3,2,1}, priors-per-cell {3,6,6,6,6,6}, 300x300, per-class sigmoid.
  • SSD-Mobile-320: feats {20,10,5,3,2,1}, priors-per-cell {3,6,6,6,6,6}, 320x320, per-class sigmoid.
  • SSDlite-Mobile-320: feats {20,10,5,3,2,1}, priors-per-cell {6,6,6,6,6,6}, 320x320, softmax. All use grouped per-level localization heads (depth = 4 * priors-per-cell) paired with class-confidence heads (depth = num_classes * priors-per-cell), and require a stretch preprocessing resize; any other head set, model frame or resize is rejected.

SSD num_classes is always derived from the confidence-head depth. The SSD300 recipe permits a contiguous prefix selection; both SSD-Mobile recipes and SSDlite-Mobile-320 require the exact encoded class count.

Definition at line 53 of file BoxDecodeType.h.

BoxDecodeTypeOption

enum class simaai::neat::BoxDecodeTypeOption : std::int32_t

strong

Tensor packing/layout option within a decode family.

Enumeration values

Auto	Backend infers the layout from tensor shapes (= 0)
PackedPerHead	Each head holds a single packed tensor (box+obj+cls) (= 1)
InterleavedByHead	Heads interleaved within tensors (= 2)
GroupedByRole	Tensors grouped by role (box, score, class) (= 3)
Split3Interleaved	Three split tensors, head-interleaved (= 4)
Split3Grouped	Three split tensors, grouped by role (= 5)
InterleavedByHeadProbability	Interleaved-by-head, class scores as probabilities (= 6)
InterleavedByHeadLogit	Interleaved-by-head, class scores as logits (= 7)
GroupedByRoleProbability	Grouped-by-role, class scores as probabilities (= 8)
GroupedByRoleLogit	Grouped-by-role, class scores as logits (= 9)

Some families admit multiple equivalent head layouts (packed vs interleaved, grouped by role, probability vs logit class scores). Auto lets the planner pick from observed tensor geometry; the explicit values force a particular decoding contract for ambiguous models.

Definition at line 92 of file BoxDecodeType.h.

FormatTag

enum class simaai::neat::FormatTag

strong

Identifies a media or tensor payload format.

Enumeration values

Auto	Unset / framework decides (= 0)
RGB	Packed RGB, 8 bits per channel
BGR	Packed BGR, 8 bits per channel (OpenCV default)
GRAY8	Single-plane 8-bit grayscale
NV12	YUV 4:2:0, Y plane + interleaved UV plane
I420	YUV 4:2:0, three planes (Y, U, V)
YUYV	YUV 4:2:2 packed (Y0 U Y1 V)
ENCODED	Generic encoded payload (codec from caps)
H264	H.264 access unit / NAL stream
AVC	Alias for H.264 (= H264)
H265	H.265 / HEVC access unit / NAL stream
HEVC	Alias for H.265 (= H265)
ByteStream	Opaque byte stream; downstream interprets bytes by contract
MLA	MLA-tessellated tensor payload
BBOX	Decoded bounding-box byte stream
ARGMAX	Argmax/segmentation map
DETESSDEQUANT	Detessellated + dequantized tensor payload
FP32	IEEE-754 32-bit float tensor
INT8	Signed 8-bit integer tensor
UINT8	Unsigned 8-bit integer tensor
BF16	bfloat16 tensor
EVXX_FLOAT32	EV-side alias for FP32
EVXX_INT8	EV-side alias for INT8
EVXX_BFLOAT16	EV-side alias for BF16

Used in caps strings and option fields throughout the pipeline. Auto is the unset sentinel (let the framework pick or sniff). The EVXX_ variants are aliases preferred by the EV74 caps surface. AVC and HEVC are aliases for H264 and H265, matching the codec spellings accepted by RtspCodec and SimaDecodeType; H264 and H265 stay the canonical serialized names.

See Also

FormatSpec

Definition at line 38 of file FormatSpec.h.

OutputMemory

enum class simaai::neat::OutputMemory

strong

How output Tensors relate to the underlying GStreamer buffers.

Enumeration values

Auto	(= 0)
ZeroCopy
Owned

Auto lets the framework pick based on platform and pipeline shape. ZeroCopy shares storage with GStreamer (faster but lifetime-coupled to the Run). Owned copies into a framework-owned buffer (safer, slightly slower).

Definition at line 104 of file Run.h.

OverflowPolicy

enum class simaai::neat::OverflowPolicy

strong

What push() does when the input queue is full.

Enumeration values

Block	push() blocks until queue space frees up. Lossless. Use for batch processing (= 0)
KeepLatest	Drop the oldest queued frame to make room
DropIncoming	Drop the new frame; keep what's queued

The right choice depends on the input source — file batches want lossless, live cameras want freshness, network feeds with chokepoint pipelines want bounded memory.

Definition at line 74 of file Run.h.

PullStatus

enum class simaai::neat::PullStatus

strong

Result status of Run::pull().

Enumeration values

Ok	A sample is available in the output parameter
Timeout	The wait elapsed without a sample arriving
Closed	The pipeline has reached EOS; no more samples will come
Error	A runtime error occurred; check the optional PullError

Definition at line 476 of file GraphOptions.h.

RunMode

enum class simaai::neat::RunMode

strong

Internal timing mode a Run operates in.

Enumeration values

Async	Continuous pipeline; user pushes/pulls asynchronously
Sync	One frame in, one result out, synchronously

Public Graph users should not pass this into Graph::build(...): use Graph::build(...) for reusable push/pull runners and Graph::run(...) for one-shot execution. The runtime still keeps this enum to select internal optimized paths.

See Also

Graph::build, Graph::run

Definition at line 232 of file GraphOptions.h.

RunPreset

enum class simaai::neat::RunPreset

strong

Convenience preset bundles for RunOptions.

Enumeration values

Realtime	Low-latency; small queues; KeepLatest overflow
Balanced	Default; moderate queues; Block overflow
Reliable	Lossless; deeper queues; Block overflow

Each preset adjusts queue depth, overflow policy, and output ownership to a profile that's known to work well for one workload class. Use start_measurement() when you want performance data for any preset.

Definition at line 90 of file Run.h.

SampleKind

enum class simaai::neat::SampleKind

strong

What kind of payload a Sample carries.

Enumeration values

Tensor	Single tensor payload (the tensor field is set)
TensorSet	Multiple tensors at one logical output index (the tensors field is set)
Bundle	Recursive: payload is a vector of Samples (the fields field is set)
Unknown	Default/uninitialized

The framework's outputs come in three shapes depending on the model's output topology: a single Tensor, a flat list of Tensors (TensorSet), or a recursive Bundle of Samples (Bundle, used by multi-logical-output models).

See Also

Sample

Definition at line 465 of file GraphOptions.h.

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.